Selected work

Representative Papers

Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought

Apr 08, 2025

This work addresses the challenge of efficiently scaling large language models (LLMs) to multimodal vision-language understanding while preserving strong textual reasoning capabilities. We propose a lightweight multimodal reasoning framework that avoids retraining language or vision backbones, instead leveraging a learnable visual projector for text-image joint reasoning. To enhance cross-modal alignment, we innovatively integrate supervised fine-tuning (SFT) with group-relative policy optimization (GRPO). Furthermore, we introduce adaptive-length chain-of-thought distillation, dynamically optimizing reasoning chain length to balance inference efficiency and accuracy. Evaluated on benchmark suites, our 38B-parameter model achieves 69.0 on MMMU and 67.5 on MathVista, while maintaining state-of-the-art textual reasoning performance (72.0 on AIME and 94.0 on MATH500). All model weights are publicly released.

1 citationsRead paper

Optimization Hyper-parameter Laws for Large Language Models

Sep 07, 2024arXiv.org

Efficient selection of dynamic hyperparameters—such as learning rate—remains challenging in large language model (LLM) training. Method: This paper proposes Opt-Laws, a novel framework grounded in stochastic differential equations (SDEs), which establishes the first interpretable mathematical relationship between hyperparameter dynamics and training loss. Opt-Laws enables *a priori* prediction of optimal learning rate schedules across pretraining, continued training, and fine-tuning. It integrates SDE-based modeling, hyperparameter–loss function fitting, and multi-scale empirical validation. Contribution/Results: Evaluated across diverse model scales and dataset sizes, Opt-Laws achieves high-accuracy training loss prediction. Experiments demonstrate that it substantially reduces hyperparameter search overhead, shortens tuning cycles by multiple-fold, and improves final model performance.

1 citationsRead paper

Two-Stage Reinforcement Learning for Sound and Adversarial Test Generation in Code LLMs

Sep 03, 2026

本文通过提出一个两阶段强化学习框架TCS来自动生成有效测试用例，以解决代码生成中高质量测试用例稀缺的问题。

0 citationsRead paper

HarMoE: Multi-Source Chest Radiograph Pretraining with Dataset-Disentangled Experts

Aug 03, 2026

This work addresses the limitations of existing chest X-ray vision-language models, which predominantly rely on a single data source (e.g., MIMIC-CXR) and thus fail to leverage the richer, more explicit disease supervision signals available in multi-label classification datasets, often conflating clinical semantics with dataset-specific artifacts. To overcome this, the authors propose HarMoE, a novel framework that employs a dataset-aware mixture-of-experts (MoE) mechanism within a deep decoder to disentangle shared medical semantics from dataset-specific features. HarMoE further introduces masked multi-dataset supervised training based on a unified disease vocabulary. This approach enables, for the first time, effective knowledge fusion across heterogeneous chest X-ray datasets, significantly outperforming strong baselines in zero-shot classification, out-of-distribution generalization, and lesion localization tasks.

0 citationsRead paper

Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents

Jun 03, 2026

This work addresses the limitation of existing evaluations for large language model (LLM) agents, which predominantly rely on end-to-end success rates and fail to diagnose root causes of planning failures. To this end, we introduce APB—the first fine-grained diagnostic benchmark specifically designed to assess LLM agents’ planning capabilities—encompassing 22 domains, 5 experimental settings, and 4,209 multimodal tasks. APB systematically evaluates critical planning skills, including goal decomposition, tool selection, constraint reasoning, and recognition of infeasible tasks. It features a hierarchical planning evaluation protocol, externally validated via ToolSandbox and τ²-bench, enabling comprehensive assessment of long-horizon planning, robustness to distractions, calibrated rejection mechanisms, and inference-time optimization. Experiments across 12 multimodal LLMs reveal widespread planning deficiencies, while APB-guided refinements substantially improve plan correctness, scoring, and downstream execution performance in three representative models.

0 citationsRead paper

Recent publications

Latest Papers

Two-Stage Reinforcement Learning for Sound and Adversarial Test Generation in Code LLMs

Sep 03, 2026

本文通过提出一个两阶段强化学习框架TCS来自动生成有效测试用例，以解决代码生成中高质量测试用例稀缺的问题。

0 citationsRead paper

HarMoE: Multi-Source Chest Radiograph Pretraining with Dataset-Disentangled Experts

Aug 03, 2026

This work addresses the limitations of existing chest X-ray vision-language models, which predominantly rely on a single data source (e.g., MIMIC-CXR) and thus fail to leverage the richer, more explicit disease supervision signals available in multi-label classification datasets, often conflating clinical semantics with dataset-specific artifacts. To overcome this, the authors propose HarMoE, a novel framework that employs a dataset-aware mixture-of-experts (MoE) mechanism within a deep decoder to disentangle shared medical semantics from dataset-specific features. HarMoE further introduces masked multi-dataset supervised training based on a unified disease vocabulary. This approach enables, for the first time, effective knowledge fusion across heterogeneous chest X-ray datasets, significantly outperforming strong baselines in zero-shot classification, out-of-distribution generalization, and lesion localization tasks.

0 citationsRead paper

Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents

Jun 03, 2026

This work addresses the limitation of existing evaluations for large language model (LLM) agents, which predominantly rely on end-to-end success rates and fail to diagnose root causes of planning failures. To this end, we introduce APB—the first fine-grained diagnostic benchmark specifically designed to assess LLM agents’ planning capabilities—encompassing 22 domains, 5 experimental settings, and 4,209 multimodal tasks. APB systematically evaluates critical planning skills, including goal decomposition, tool selection, constraint reasoning, and recognition of infeasible tasks. It features a hierarchical planning evaluation protocol, externally validated via ToolSandbox and τ²-bench, enabling comprehensive assessment of long-horizon planning, robustness to distractions, calibrated rejection mechanisms, and inference-time optimization. Experiments across 12 multimodal LLMs reveal widespread planning deficiencies, while APB-guided refinements substantially improve plan correctness, scoring, and downstream execution performance in three representative models.

0 citationsRead paper

How LLMs Are Persuaded: A Few Attention Heads, Rerouted

May 10, 2026

Large language models are susceptible to persuasion that leads them to deviate from factual responses, yet the underlying mechanisms remain poorly understood. This work addresses this gap by employing attention head interventions, latent space geometric analysis, and feature manipulation to reveal, for the first time, that the persuasion process is governed by a small subset of attention heads in intermediate layers, forming a narrow and monitorable causal circuit. The study identifies a one-dimensional critical feature responsible for evidence routing, along with its upstream generative source. Experiments across multiple open-source models and real-world adversarial scenarios demonstrate that directly modifying or removing this feature effectively steers or blocks persuasive behavior, offering a novel pathway toward enhancing model robustness.

0 citationsRead paper

KEPIL: Knowledge-Enhanced Prompt-Image Learning for Prompt-Robust Disease Detection

May 09, 2026

Existing medical vision-language models exhibit high sensitivity to prompt variations and lack reliable integration of external knowledge, hindering their clinical deployment. To address this, this work proposes a prompt optimization framework that synergistically combines structured medical ontologies with large language models. The approach introduces three key mechanisms: dynamic prompt augmentation, semantic-aware contrastive loss, and entity-centric report standardization, which collectively enhance prompt robustness and zero-shot generalization in radiology disease detection. Evaluated across seven benchmarks, the method achieves state-of-the-art performance, demonstrating significant improvements under prompt perturbations—most notably a 6.37% increase in AUC on the CheXpert dataset and an average gain of 4.11% across datasets.

0 citationsRead paper